Decoding AI: How “Excess Words” Reveal the Hidden Footprint of Generative AI in Scientific Writing

To go back to this article, head to My Profile and then click on View saved stories.

Identifying AI-Generated Text: Uncovering the Clues

To date, even the creators of artificial intelligence have struggled to develop effective strategies for pinpointing texts crafted by expansive language algorithms. However, a team of researchers has now devised an innovative approach to gauge the prevalence of large language model (LLM) utilization within a broad corpus of academic texts. They've done this by tracking the surge in usage of certain "superfluous words" that became notably more common in the period identified as the LLM era, specifically the years 2023 and 2024. The findings from this study indicate that "a minimum of 10 percent of the abstracts from 2024 underwent processing via LLMs," the research team reports.

In a preliminary research document shared this month, a team of four scholars from the University of Tübingen in Germany and Northwestern University in the United States revealed their motivation came from research that evaluated the effects of the Covid-19 pandemic by examining the surplus deaths against recent historical data. By adopting a comparable approach to assess the "surplus in word usage" following the widespread adoption of LLM (large language models) writing aids in late 2022, the team observed a sudden surge in the usage of specific stylistic words, a phenomenon they described as unparalleled in both its nature and scale.

Exploring the Topic

The study involved examining the shifts in vocabulary by scrutinizing 14 million abstracts from papers listed on PubMed, spanning from 2010 to 2024. This was done by monitoring how often each word showed up year over year. The researchers then matched the predicted usage rates of these words (which were projected from trends before 2023) against their real usage rates in the years 2023 and 2024, a period marked by the extensive utilization of LLMs.

This article was first published on Ars Technica, a reliable platform for updates on technology, analysis of tech policies, critiques, among other content. Ars Technica is a subsidiary of Condé Nast, the same corporation that owns WIRED.

The investigation revealed a series of terms that were relatively rare in scientific summaries before 2023, which then experienced a significant rise in occurrence following the introduction of LLMs. For example, the term "delves" was mentioned in 2024 documents 25 times more than what was anticipated based on trends prior to LLMs; similarly, the usage of terms such as "showcasing" and "underscores" saw a ninefold increase. Additionally, words that were already common in these abstracts saw an uptick in their frequency after LLMs came into play: the term "potential" saw an increase of 4.1 percentage points, "findings" rose by 2.7 percentage points, and "crucial" went up by 2.6 percentage points.

Alterations in the way words are utilized can occur without the involvement of large language models (LLMs)—it's simply a part of how language naturally evolves, with certain terms becoming more or less popular over time. Nevertheless, the study highlighted that, before the advent of LLMs, such rapid and significant yearly increases in the usage of specific words were typically associated with significant global health crises: for instance, "ebola" surged in popularity in 2015; "zika" in 2017; and terms such as "coronavirus," "lockdown," and "pandemic" experienced a spike from 2020 to 2022.

During the era following the introduction of Large Language Models (LLMs), researchers identified numerous words that experienced a sharp rise in usage within scientific literature, unrelated to global happenings. Unlike the spike in noun usage linked to the Covid pandemic, this period saw a dominant increase in the use of "style words" such as verbs, adjectives, and adverbs. Examples of these words include "across, additionally, comprehensive, crucial, enhancing, exhibited, insights, notably, particularly, within".

The observation that the term "delve" is appearing more frequently in scientific literature isn't groundbreaking—it's something that has been recognized before, particularly in recent times. However, earlier research typically depended on contrasting these findings with authentic human-written texts or with sets of indicators specific to large language models (LLMs) that were identified externally from the research at hand. In this instance, the collection of abstracts from before 2023 serves as a comparative baseline, effectively illustrating the shift in word usage in the scientific community following the widespread adoption of LLMs.

A Complex Interaction

Researchers have pointed out the increased frequency of certain "indicator words" in the era following the introduction of large language models (LLMs), making it somewhat straightforward to identify instances of LLM application. Consider the following example of an abstract sentence highlighted by the study, with the indicator words emphasized: "An in-depth understanding of the complex interaction among […] and […] is crucial for successful treatment approaches."

Following an analysis of the frequency of specific keywords within single studies, the research team suggests that a minimum of 10 percent of the academic articles published after 2022 in the PubMed database likely had some form of assistance from large language models (LLMs). The actual figure could surpass this estimate, according to the researchers, as their methodology might not capture all instances of LLM-supported abstracts that lack the keywords they were tracking.

The study revealed significant variations in the observed percentages among various groups of papers. It was noted that research papers from countries such as China, South Korea, and Taiwan exhibited markers indicative of Large Language Model (LLM) contributions about 15 percent of the time. This finding leads to the speculation that "LLMs might assist non-native English speakers in refining their English manuscripts, potentially explaining their widespread adoption." Conversely, the researchers propose that native English speakers "could be more adept at identifying and eliminating awkwardly phrased words produced by LLMs," thereby concealing their use of LLMs from this type of scrutiny.

Identifying the employment of Large Language Models (LLMs) is crucial, the scholars emphasize, due to the notorious tendency of LLMs to fabricate references, deliver erroneous summaries, and assert unfounded claims that appear credible and persuasive. However, as awareness of the specific indicator words associated with LLMs becomes more widespread, human editors might improve at removing these words from the produced text prior to its distribution globally.

It's conceivable that, in time, advanced language models could perform their own analysis of word usage patterns, adjusting the significance of certain keywords to make their responses appear more naturally human. Soon enough, we might find ourselves in a scenario where we require the expertise of Blade Runners to identify the generative AI content camouflaged among us.

Originally published on Ars Technica, this story has been shared here

Suggested for You …

Direct to your email: Fast Forward by Will Knight delves into the latest progress in artificial intelligence.

Delving into the largest undercover operation ever conducted by the FBI

The WIRED AI Elections Initiative: Monitoring over 60 worldwide polls

Ecuador finds itself utterly without electricity due to a severe drought.

Be confident: Here's a list of the top mattresses available for online purchase.

Additional Coverage from WIRED

Evaluations and Instructions

© 2024 Condé Nast. All rights are protected. WIRED could receive a share of revenue from the sale of products linked on our website, a result of our Affiliate Agreements with retail partners. Content from this site is not allowed to be copied, shared, broadcast, stored, or used in any form without explicit written consent from Condé Nast. Advertising Options

Choose a global website